Arabic & Multilingual Aspect-Based Sentiment Analysis with MARBERT

A Natural Language Processing project for aspect-based sentiment analysis (ABSA) on Arabic and multilingual customer reviews.

The model detects which business aspects are mentioned in a review, then predicts the sentiment for each detected aspect.

Supported aspects include:

• Food
• Service
• Price
• Cleanliness
• Delivery
• Ambiance
• App experience
• General opinion

The project uses MARBERT because it is well suited for Arabic, dialectal Arabic, and informal social-media-style text.

---

Project Overview

Traditional sentiment analysis usually predicts one overall label for a full review.

This project goes deeper by answering questions such as:

The customer liked the food, but did they complain about the price?
Was the delivery experience negative even though the app experience was positive?

For each review, the model returns a structured JSON output like:

{
  "review_id": "example_001",
  "aspects": ["food", "service", "price"],
  "aspect_sentiments": {
    "food": "positive",
    "service": "negative",
    "price": "negative"
  }
}

---

Key Features

• Arabic and dialectal Arabic support
• Mixed Arabic-English review handling
• Egyptian Arabic and informal review examples
• Aspect detection and sentiment classification
• JSON output for dashboards, APIs, or competition submissions
• Gradio demo interface for public presentation
• Kaggle-friendly training workflow

---

Model Approach

The task is transformed into a pairwise classification problem.

For every review, the model checks each possible aspect separately.

Example:

Review:
الخدمة ممتازة والأكل سيء

Aspect:
food

The model predicts one of the following labels:

• not_mentioned
• positive
• negative
• neutral

This allows the system to identify both:

1. Whether an aspect is mentioned
2. The sentiment toward that aspect

---

Repository Structure

arabic-multilingual-absa-marbert/
│
├── README.md
├── requirements.txt
├── multilingual-nlp-marbert.ipynb
├── gradio_absa_demo_app.py
├── demo_examples.md
└── .gitignore

Recommended optional folders:

images/
outputs/

Do not upload large trained model files or full Kaggle datasets directly to GitHub.

---

Dataset

The project was developed on Kaggle using the DeepX NLP dataset.

Expected Kaggle dataset path:

DATA_DIR = "/kaggle/input/datasets/kareemmaged/deepx-nlp"

Expected files:

train_fixed(Sheet1).csv
validation_fixed(Sheet1).csv
unlabeled_fixed(Sheet1).csv
DeepX_hidden_test (4).xlsx

The training and validation files contain labeled aspect and sentiment columns.

The unlabeled and hidden test files contain review text and metadata, and the trained model generates the predictions.

---

How to Run on Kaggle

1. Open the notebook

Upload or open:

multilingual-nlp-marbert.ipynb

2. Enable Kaggle settings

In the Kaggle notebook settings:

Accelerator: GPU
Internet: On

Internet is needed for downloading the pretrained MARBERT model and for launching a public Gradio demo link.

3. Install dependencies

The notebook includes an installation cell, but you can also install dependencies manually:

pip install -r requirements.txt

4. Check dataset paths

Inside the notebook, make sure the paths match your Kaggle dataset location:

DATA_DIR = "/kaggle/input/datasets/kareemmaged/deepx-nlp"

TRAIN_FILE = f"{DATA_DIR}/train_fixed(Sheet1).csv"
TEST_FILE = f"{DATA_DIR}/validation_fixed(Sheet1).csv"

For final prediction, change TEST_FILE to:

TEST_FILE = f"{DATA_DIR}/unlabeled_fixed(Sheet1).csv"

or:

TEST_FILE = f"{DATA_DIR}/DeepX_hidden_test (4).xlsx"

5. Run all notebook cells

The notebook will:

1. Load the dataset
2. Preprocess review text
3. Convert reviews into aspect-level training rows
4. Fine-tune MARBERT
5. Evaluate on validation data
6. Tune the aspect detection threshold
7. Save the best model and prediction artifacts

The trained model is saved to:

/kaggle/working/marbert_absa_output/best_model

The metadata/artifacts file is saved to:

/kaggle/working/marbert_absa_output/artifacts.json

---

Gradio Demo

The repository includes a Gradio interface:

gradio_absa_demo_app.py

The demo allows users to paste a review and see:

• Detected aspects
• Sentiment for each aspect
• Confidence scores
• Final JSON prediction
• All-aspect confidence table

To run it on Kaggle after training:

python gradio_absa_demo_app.py

The app should launch using:

demo.queue().launch(share=True, debug=True)

share=True creates a public demo link that can be used for recording or sharing.

---

Demo Examples

Arabic Review

الخدمة كانت محترمة جدا والمكان رايق، بس الأكل وصل بارد واللحمة ناشفة والسعر مبالغ فيه

Expected behavior:

• Service: positive
• Ambiance: positive
• Food: negative
• Price: negative

Mixed Arabic and English

The app is smooth and checkout was fast, but el delivery اتأخر ساعتين والدعم مردش عليا خالص.

Expected behavior:

• App experience: positive or neutral
• Delivery: negative
• Service: negative

Egyptian Arabic Franko

el akl kan gamed awy bas el service slow w el bill kan aghla mn el expected.

Expected behavior:

• Food: positive
• Service: negative
• Price: negative

---

Output Format

The model returns predictions in this format:

{
  "review_id": "review_001",
  "aspects": ["food", "service"],
  "aspect_sentiments": {
    "food": "positive",
    "service": "negative"
  }
}

If no aspect is detected, the output becomes:

{
  "review_id": "review_001",
  "aspects": ["none"],
  "aspect_sentiments": {}
}

---

Notes and Limitations

• The model is strongest on Arabic-script and dialectal Arabic reviews.
• Franco-Arabic examples may work, but performance depends on how much similar text exists in the training data.
• Very short or sarcastic reviews can be harder to classify.
• The detection threshold affects how many aspects are included in the final output.
• The trained model files are not included in this repository because they can be large.

---

Future Improvements

Possible next steps:

• Add more labeled Franco-Arabic examples
• Add model comparison with AraBERT or XLM-R
• Deploy the Gradio demo on Hugging Face Spaces
• Add explainability for aspect predictions
• Add a dashboard for analyzing review trends by business or category

---

Author

Created by Kareem Maged as part of a multilingual NLP project focused on Arabic customer review understanding.